Big fan of this, but I still feel like getting anything working through ROCM is a chore and a half. There's a dire need for standardisation across hardware and software stacks in ML.